Claude Fable 5 Mythos 5
Anthropic announces new Claude models, Fable 5 and Mythos 5.
9 articles · 4 categories
What happened in AI — Jun 9, 2026
Tuesday, Jun 9, 2026
In 30 seconds
- Anthropic announced new Claude models, Fable 5 and Mythos 5.
- OpenAI published two Codex customer stories — Nextdoor (with GPT-5.5) and Notion — on real-world coding-agent workflows.
- Latent Space launched FrontierCode, a benchmark for code quality over "slop."
- Open-source search agent Harness-1 reportedly outperforms GPT-5.4 on recall.
- Cost control surfaced as a theme: a local TypeScript guardrail for agent cost failures and tooling to track per-model token spend.
- Karpathy on software "on a tap": as it gets cheaper to produce, demand for it grows (Jevons paradox).
Anthropic set the tone on the frontier, announcing new Claude models — Fable 5 and Mythos 5 — the day's headline release. Around it, the story was overwhelmingly about coding agents proving their worth in production: OpenAI published back-to-back customer dispatches on how Nextdoor and Notion are building with Codex, from chasing hard-to-reproduce bugs to one-shotting specs and shipping features across small teams.
Underneath the productivity gains ran a quieter thread about quality and cost. Latent Space introduced FrontierCode, a benchmark explicitly aimed at "code quality over slop"; a developer shipped a local TypeScript guardrail to cap AI agent cost failures; and an open-source search agent, Harness-1, claimed to beat GPT-5.4 on recall. Andrej Karpathy framed the moment well — as working software "comes out on a tap," Jevons paradox kicks in and demand for software only grows.
Tooling rounded out the day: AWS walked through an agentic incident-triage assistant, and Simon Willison shared a workflow for tracking per-model token costs across the coding agents running on his laptop.
Frontier Models 1 item
The day's marquee release from a frontier lab.
Coding Agents at Work 3 items
Real-world dispatches on teams putting coding and operations agents into production workflows.
How engineers at Nextdoor use Codex to build without limits
Nextdoor's engineers use Codex with GPT-5.5 to investigate hard-to-reproduce issues, build across platforms, and stay focused on product outcomes.
What Codex unlocks for Notion
Notion uses Codex to one-shot specs, build AI Voice Input for the web, and multiply engineering output across small teams.
Build an agentic incident triage assistant with Amazon Quick and New Relic
An AWS walkthrough for building a custom incident-triage agent with Amazon Quick and New Relic, applied to one of engineering's most time-sensitive workflows.
Open Source & Agent Tooling 3 items
Open models and the practical scaffolding around agents — recall-focused search, cost guardrails, and usage visibility.
Open Source Agent, Harness-1, Outperforms GPT-5.4 on Recall
Researchers trained an open-source AI search agent, Harness-1, that reportedly beats GPT-5.4 at recalling relevant information.
I built a local TypeScript guardrail for AI agent cost failures
ai-costguard, a local TypeScript guardrail aimed at catching and capping runaway AI agent cost failures.
Setting a custom price for a model in AgentsView
Simon Willison on using AgentsView (by Wes McKinney) to explore token usage across coding agents on his laptop, including setting custom per-model pricing.
Benchmarks & Commentary 2 items
Measuring code quality rather than volume — and reflecting on what abundant software means.
[AINews] FrontierCode: Benchmarking for Code Quality over Slop
Latent Space introduces FrontierCode, a benchmark designed to measure code quality rather than reward high-volume "slop."
Quoting Andrej Karpathy
Karpathy on software increasingly "coming out on a tap": as it gets cheaper to produce, Jevons paradox kicks in and demand for software grows substantially.